Whisper Small Hy - Erik Mkrtchyan

This model is a fine-tuned version of openai/whisper-small on the Hy Generated Audio Data with CV 20.0 dataset. It achieves the following results on the evaluation set:

• Loss: 0.1185
• Wer: 26.9720

Model description

This model is based on OpenAI's Whisper Small and fine-tuned for Armenian using a combination of real and synthetic audio data. It is designed to transcribe Armenian speech into text.

Intended uses & limitations

Intended Uses:

• Armenian speech-to-text applications
• Research on ASR for low-resource languages
• Educational and experimental projects involving Whisper models

Limitations:

• May not generalize well to accents or noisy audio not represented in the training set
• he model may hallucinate text or produce inaccurate transcriptions, especially on unusual or out-of-distribution inputs, due to the inclusion of TTS-generated synthetic data in training.

Training and evaluation data

The dataset contains both real and high-quality synthetic Armenian speech clips.

Total duration: ~145 hours
Train set duration(train+generated): ~127 hours
Test set duration(test+eval) ~18 hours

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 16
• eval_batch_size: 8
• seed: 42
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 15000
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.51.3
• Pytorch 2.7.0+cu126
• Datasets 3.6.0
• Tokenizers 0.21.1